saphes

Readability and lexical diversity — two metrics, done carefully, with the parameters other implementations hardcode.

saphes — σαφής, "clear, plain, distinct". Aristotle makes clarity the chief virtue of λέξις (style); the other classical axis is ποικιλία, variety. The two metrics here are exactly those axes: LIX measures clarity, TTR measures variety.

Why this exists

textstat, textdescriptives, lexicalrichness and taaled already cover this ground. Two reasons to still build it:

1. The LIX long-word threshold is hardcoded at 6 everywhere. That 6 comes from Björnsson's Swedish original, and it is wrong for the languages we work on. Hungarian is agglutinative and Ancient Greek heavily inflected, so at threshold 6 nearly every token counts as "long" and the index saturates into a flat line. Measured over the full Hungarian Webcorpus, 44.5% of running tokens are "long" at threshold 6, against 25.7% in Swedish. On real Hungarian prose that pushes LIX to 60.4 — "very difficult" — where the calibrated threshold gives 43.4. Parameterising the threshold is the whole point.
2. Implementations disagree. They count words, sentences and long words differently, so they rank the same texts differently. So: expose the counts, document every choice, make results auditable.

Non-goal: becoming another kitchen-sink readability library.

Installation

uv add saphes

The core has no dependencies — plain Python and the standard library.

Quickstart

from saphes import lix, lexical_diversity

# LIX takes SURFACE FORMS. Word length is the signal.
result = lix("The cat sat on it. Complicated sentences generally frighten us.")
result.score            # 45.0
result.words, result.sentences, result.long_words   # (10, 2, 4)  -> A, B, C
result.band             # 'standard'

# Raise the threshold for inflected languages, where 6 saturates.
hu = "A gyermekeknek megmutatták a településeken található nevezetességeket. Elutaztak."
lix(hu).score                                  # 79.0 — an ordinary sentence, "very difficult"
lix(hu).long_word_share                        # 0.75 — three words in four are "long"
lix(hu, long_word_threshold=9).score           # 54.0 — discrimination restored
lix(hu, long_word_threshold=9).long_word_share # 0.50

# Or use the calibrated default, with its provenance attached.
from saphes import recommended_threshold
rec = recommended_threshold("hu")
rec.threshold        # 8
rec.matched_share    # 0.273 — against Swedish 0.257 at threshold 6
lix(hu, long_word_threshold=int(rec))

# Diversity takes LEMMAS, and `unit` is required — no default.
lexical_diversity(["ház", "kutya", "ház"], unit="lemma")

# Comparing texts of different lengths? Use MATTR, not TTR.
lexical_diversity(lemmas, unit="lemma", window=100).mattr

The data contract

The single most important thing in this package: the two metrics require opposite token streams.

Metric	Required input	Why
lexical_diversity	lemmas	Surface variation is noise — it measures morphology, not vocabulary. Hungarian ház / házak / házban / házakat is four types and one lemma.
lix	surface forms	Word length is the signal. házakban is 8 characters; its lemma ház is 3. Lemmatising erases the measurement.

Feed the same list to both and exactly one of them is silently wrong — no error, no NaN, just a plausible number. Four guards exist against that:

• unit is a required keyword on lexical_diversity, with no default.
• The parameter names differ: lemmas= versus words=. Crossing them raises.
• A raw string is refused by lexical_diversity unless you pass unit="surface", since a string can only ever yield surface forms.
• Every result records the unit it measured, so any serialised table says which stream produced each number.

saphes consumes lemmas; it does not produce them. Lemmatisation is language-specific and heavy — CLTK or a treebank for Greek, huspacy for Hungarian. The caller lemmatises; saphes measures.

Features

• lix(...) — the readability index, with long_word_threshold as a first-class parameter. A long word is len(word) > threshold, so the default 6 means seven letters or more, per Björnsson's "more than six letters".
• Three ways to supply the sentence count B — segmented from raw text with a pluggable splitter, pre-split, or an explicit integer. Neither a treebank that drops punctuation nor a spaCy pipeline with the parser disabled can give you sentences, so the explicit path is not a corner case. The result records which was used.
• lexical_diversity(...) — TTR and MATTR, with the token unit declared and recorded.
• mattr(tokens, window=100) — the length-corrected moving average, as a bare float, for drop-in use.
• Calibrated thresholds, with their provenance. recommended_threshold("hu") returns the empirically matched threshold and the shares behind it, the runner-up, which of six independently computed curves agreed, and the caveats. The number ships; the corpus does not. The study is in experiments/lix_calibration/.
• Counts, not just scores. Every function returns A, B, C (or types and tokens) plus every parameter used and the saphes version. A bare float is unauditable.
• Explicit length policy. Decomposed Unicode otherwise inflates every word length, which hits polytonic Greek and accented Hungarian hardest. length_policy defaults to NFC normalisation and can be swapped for grapheme counting, raw code points, or your own callable.

Warning

TTR is inversely related to text length. TTR values from texts of different lengths are not comparable — a raw TTR over corpora of different sizes mostly ranks them by size. This matters directly for per-decade and per-book work, where lengths always differ. Pass window= and read .mattr instead.

Documentation

Full docs at saphes.readthedocs.io; sources in docs/. Worked examples live in examples/ and are included in the docs verbatim, so they cannot drift.

Roadmap

• LIX with a parameterised long-word threshold
• TTR and MATTR with a required, recorded token unit
• RIX (long words per sentence)
• Empirically calibrated per-language thresholds, from token-weighted word-length distributions — Hungarian ships as recommended_threshold("hu")
• The same study for Ancient Greek, for the Homer project
• POS-filtered diversity, once lemmas carry tags
• MTLD, HD-D, vocd-D, Maas

Maintenance

• Logo and README banner
• Mutation-testing baseline

Explicitly out of scope: Flesch, Kincaid, SMOG and relatives. They need syllabification, which is language-specific and a different project. LIX was chosen precisely because it needs only word length and sentence count, so it travels across languages.

Made by

saphes is made by Crow Intelligence.

License

MIT